The Gold Standard of Experimental Design | Surrogate Endpoints, Causal Mediation and Outcome Substitution

1. Name the outcome people actually care about

Back to contents

Begin with the clinical, behavioural or functional outcome the decision is meant to improve: survival, symptoms, mobility, learning or quality of life. A laboratory value, image or intermediate event may be informative, but it is not the benefit merely because it is measurable.

Write the target outcome, time horizon and decision context before proposing a surrogate. This prevents convenience from silently redefining success.

2. Separate biomarker, mediator and surrogate

Back to contents

A biomarker indicates a biological or other process. A mediator lies on a causal pathway from treatment to outcome under stated assumptions. A surrogate endpoint is intended to substitute for the meaningful outcome when estimating treatment benefit. One measure can occupy different roles in different settings.

Correlation with outcome risk does not establish surrogacy. A prognostic marker may identify high-risk people while failing to predict how changing the marker changes the outcome.

3. Draw the pathways before fitting the model

Back to contents

Represent treatment T, surrogate S and outcome Y. The hoped-for pathway is T→S→Y. But treatment may also affect Y through another path, create harm not captured by S, or alter S without changing the process that determines Y. Common causes can associate S and Y without making S a valid substitute.

Use the diagram to specify estimands and assumptions, not as decoration.

4. Match validation strength to the intended use

Back to contents

A candidate surrogate is still under evaluation. A reasonably likely surrogate has strong mechanistic or epidemiologic support but incomplete clinical validation. A validated surrogate has strong evidence that an effect on it predicts a specific clinical benefit in a defined context.

Validation is not a permanent badge. It depends on treatment class, population, disease stage, comparator and outcome. State the exact context in which substitution is proposed.

5. Test both individual-level and trial-level relationships

Back to contents

Individual-level association asks whether people with better surrogate values tend to have better outcomes. Trial-level association asks whether treatment effects on the surrogate predict treatment effects on the meaningful outcome across trials. The second is essential for substitution.

A single large trial can estimate treatment effects, but it cannot reveal how those effects co-vary across diverse interventions. Meta-analytic validation needs enough trials and mechanisms to test transport.

6. Worked example: the marker improves but the outcome does not

Back to contents

A randomised treatment improves a biomarker by ten units compared with control. The biomarker is strongly associated with later complications in untreated cohorts. Yet the meaningful outcome occurs in 10% of both arms.

The pathway analysis reveals two effects: the treatment improves the biomarker through the intended mechanism, but it also causes an off-target harm that raises outcome risk. The mediated benefit and unmeasured harm cancel. A biomarker-only trial would have declared success.

The correct conclusion is not that biomarkers are useless. It is that this marker did not capture the total treatment effect in this intervention context. Follow-up on the meaningful outcome and harms remains necessary.

7. Use causal mediation without overclaiming

Back to contents

Mediation analysis can decompose a treatment effect into pathways under assumptions about post-randomisation confounding, consistency, interference and measurement. Those assumptions are demanding because the mediator is not randomly assigned even when treatment is.

A large mediated proportion in one trial does not prove the endpoint can replace outcomes across new treatments. It describes a pathway under one intervention and population.

8. Changed-context transfer check: attendance and learning

Back to contents

Move the logic to an education trial. A tutoring programme increases attendance, and attendance predicts examination performance. It is tempting to use attendance as a surrogate for learning.

But the programme might increase attendance through rewards without improving instruction, while stress or curriculum narrowing affects learning through paths attendance does not capture. Measure mastery directly and test whether programme effects on attendance predict effects on learning across settings and interventions.

9. Recognise the surrogate paradox and transport limits

Back to contents

The surrogate paradox occurs when treatment benefits the surrogate yet harms or fails to benefit the outcome. It can arise from unmeasured pathways, competing risks or a surrogate that captures only one disease mechanism.

Do not transfer a surrogate validated for one drug class to another without evidence, use a composite proxy whose dominant component is less meaningful, or stop confirmatory follow-up because an early marker moved. Outcome substitution narrows what the experiment can establish.

10. Audit the substitution decision

Back to contents

For every surrogate-based claim, trace the meaningful outcome, proposed mechanism, validation category, population, treatment class, comparator, time horizon and evidence across trials. Prespecify the surrogate, analysis and handling of missingness.

Search for discordant trials, not only supportive associations. Reanalyse conclusions under weaker surrogate assumptions and include direct outcome and safety evidence whenever available.

11. Frequently asked questions

Back to contents

Is a biomarker a surrogate endpoint?

Only when it is used to substitute for a meaningful endpoint and has evidence appropriate to that use.

Does mediation prove surrogacy?

No. Mediation within one trial does not show that treatment effects on the mediator predict outcome effects across interventions.

Can accelerated decisions use uncertainty?

Yes, but the evidentiary category and need for confirmatory outcome data must be explicit.

12. The gold standard keeps the real outcome in view

Back to contents

Surrogate endpoints can shorten trials and expose mechanisms, but their value rests on an evidence chain: meaningful target outcome, plausible pathways, treatment-effect validation and bounded transport. A responsive proxy is a clue, not a verdict.

Before substituting, ask whether the measure predicts treatment effects, captures important harms and remains valid for this population and intervention. If those questions are unanswered, measure the outcome that matters.